Abstract
This disclosure describes techniques that enable a server-side large language model to provide personalized responses to user queries while protecting user personal data that is used in a request sent to the LLM from a user device. With user permission, entities from personal data stored locally on a user device are retrieved and hashed using a hash function at the local device. The hashed values are injected in a request sent to the LLM that is located in the cloud or other network location distinct from the local device. The LLM generates a response including one or more hashed values and sends the response back to the local device. The hashed values in the LLM response are converted back to the original entities in an output response. The described features allow personal data to be used in prompts to language models that are accessed over network connections, thus allowing personalized LLM responses to be generated without exposing user personal information.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yim, Keun Soo, "Privacy-friendly Personalization of LLM Responses Using Hashed Entity Injection", Technical Disclosure Commons, (December 19, 2023)
https://www.tdcommons.org/dpubs_series/6512